Add evaluation, minor updates to HF space

Add per-batch generation liveness logging

Upload intermediate GRPO checkpoints to HF Hub on every save

Make self-play training resilient to HF Space restarts

feat(training): improve self-play progress visibility and reward diagnostics

Minor config changes

Update training config, add checkpointing on HF

test hf space commit

Sync current main to Hugging Face Space

fix(rewards): never crash GRPO on malformed completions